Multi-Label Classification with 1-D Convolutional Neural Network

13 次查看（过去 30 天）

显示更早的评论

Jacobus Haupt 2017-6-19

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/345348-multi-label-classification-with-1-d-convolutional-neural-network

评论： Abdullah Fahim 2019-1-17

在 MATLAB Online 中打开

Hello,

I am currently running the following 1-dimensional convolutional neural network.

layers = [imageInputLayer([501 1 1])
          convolution2dLayer([25 1],20)
          reluLayer
          maxPooling2dLayer([4 1],'Stride',2)
          fullyConnectedLayer(2)
          softmaxLayer
          classificationLayer()] 
opts = trainingOptions('sgdm','MaxEpochs',2, ...
   'InitialLearnRate',0.001);
[convnet, info] = trainNetwork(data,labels,layers,opts);

However, I want some output to be able to be classified as several sorts of signals (i.e. my classes are not mutually exclusive). For example, input could be a parabola, a gaussian, or a gaussian inside a parabola (as well as "none of the above" as a class; this would be a simplified form of my problem). I know the problem is with softmax but I do not understand how to replace it with a layer of logistic sigmoids in a SeriesNetwork, if this is still what is wanted in a SeriesNetwork, if I should train with softmax then switch to logsig, etc.

Thanks